Bernal S D_2010.pdf - University of Plymouth

4.2. ARCHITECTURES
Scale band
RF size", AWs2
S2 types, Kn
Band pooling, &Sc2
Grid site, ANc2
C2 lypes, Kc7
RF si/c", A^s3
S3 types, Ks2
1
4
S2 parameters
2
3
8
12
1000
C2 parameters
All bands: 1---8
All S2 units
1000
S3 parameters
1
60
"52 protarype elements- {iN^i x AW]« x Kn (4 oriemations}. Same for all scale bands.
''S3 proiuiype elememn &Ns?i x ANxi x Iic2 (1000 Italures).
7a/)/e ^.2.- Parameters of the 3-layer archileciure. Based i»n Serre et al. l2007c).
the corresponding paramelere), and the node.s of different S2 RF sizes do not interact with each
other.
The number of nodes in each S2 set is 2253 for ANs2 - 4.1572 for ANs2 = 8,1098 for AA's: = 12
and 758 for ANsj — 16. Bigger ANsj imply less resulting S2 units as the number of S2 units is
equal to the number of CI units divided by AA'.S-T. The numberof features of each RF size, A!^ is
setlothciotalnumberof features in layer S2 divided by four i.e. K'^ — Ksil^- 1000/4 = 250.
Each node in the networli has an associated CPT which links it with its parent nodes. Similarly,
each node performs the same internal operations, shown in I'igure 4.3, which correspond to the
distributed implementation of belief propagation.
4.2.2 Alternative three-level architecture based on Yamane et at. (2006)
The parameters of this architecture are shown in Table 4.3 and the resulting Bayesian network
is iilusiraled in Figure 4.5 (only from layer S2 above, as layers SI and CI are equivalent to
the previous version). This architecture was introduced to try to improve the recognition of the
iranslnteil dataset of input objects. It is a variation of the 3-level HMAX model (Serre et al.
2007c). with a reduced pooling range (RF size) at the top layers C2 and S3, More specifically,
C2 prototypes do not pool over the whole set of S2 units, but over a smaller range (50% of the S2
map length), which leads to a 3-by-3 C2 grid of units. This then allows the S3 RF size to be set
151
4
16